Supercomputer Vision | The Juice
Zumo Labs presents The Juice, a weekly newsletter focused on computer vision problems (and sometimes just regular problems). Get it while it’s fresh.
Week of June 21–25, 2021
____
The Conference on Computer Vision and Pattern Recognition, or CVPR, wraps up tomorrow. It’s an annual opportunity for folks in the field to share their research and, in some cases, for companies like Tesla to show off. On Monday, the company’s Head of AI, Andrej Karpathy, unveiled the supercomputer they expect to power their vision-only full self-driving technology moving forward — a computer they’re saying might be the fifth-most powerful in the world.
Part of the reason Tesla needs such a beefy rig is the vast amount of data they’re capturing. Per TechCrunch, “Tesla has accumulated 1 million videos of around 10 seconds each and labeled 6 billion objects with depth, velocity and acceleration.” That’s a lot of data annotation.
All of this has us reflecting on what we’re building here at Zumo Labs, how it alleviates these pain points for folks who don’t have a supercomputer, and where it fits into the ML workflow. In our most recent blog post, our Co-Founder Kory lays it all out on the table. We’re building AutoML for Data.
____
#DataLabeling
It’s widely known that ImageNet has a label error rate of 5.8%. We’ve written previously about the problems associated with labeled data. This piece, which dives even deeper into issues like bias and the prohibitive costs of the manual labeling process, can be considered corroboration. And as expected, one of the promising paths forward the piece identifies is synthetic training data.
Hand Labeling Considered Harmful, via O’Reilly.
#WasteBot
Soft plastics like cling wrap and plastic bags can gum up the works of recycling lines, and in most cases, must be removed in a manual process. A team of researchers at the University of Sydney has designed a robot that uses computer vision to identify and pick those soft plastics off the line. The Australian Federal Government has awarded them a significant grant to pursue the work.
Recycling robot could help solve soft plastic waste crisis, via University of Sydney.
#WantBot
If you want a robot to help you around the kitchen — and let’s be honest, who doesn’t? — that robot will need to be capable of navigating both transparent and reflective surfaces. Well, this week, Toyota Research Institute proudly announced that their “robots have learned to identify any wipeable surface.” The milestone would not have been possible without synthetic data, per the company’s post: “Using synthetic data also alleviates the need for time-consuming, expensive, or impractical data collection and labeling.”
Toyota Research Institute shows how its robotics work with difficult surfaces in the home, via TechCrunch.
#GauGAN
Generative adversarial networks can be challenging to get started with. Perhaps that’s why NVIDIA liked to show off their admittedly impressive GauGAN tech at trade shows like SIGGRAPH. But as of this week, they’re finally making the app — which turns your MS Paint-grade basins into uncanny valleys — available to anyone in NVIDIA Canvas.
NVIDIA’s Canvas app turns doodles into AI-generated ‘photos’, via Engadget.
#EthicalAI
Did you know that the same algorithm used to create deepfakes is used for weather forecasting? Per DeepMind research scientist Raia Hadsell, it’s a reminder that anyone working on AI must consider the broader implications of their work. Hadsell spoke at the Lesbians Who Tech Pride Summit this week, and this piece from VentureBeat covers the beats of her call to action.
DeepMind scientist calls for ethical AI as Google faces ongoing backlash, via VentureBeat.
#RIP
Yesterday, John McAfee was found dead in the Spanish prison where he was being held, just hours after a court approved his extradition to the US. The father of commercial antivirus software and serial entrepreneur, McAfee was a deeply troubled man. This piece from 2012 is a harrowing character study and profile of the man that just so happened to coincide with a murder investigation in which he was wanted for questioning.
**Content Warning: On click, the header photo depicts McAfee holding a gun to his head.**
John McAfee Fled to Belize, But He Couldn’t Escape Himself, via Wired.
____
📄 Paper of the Week
Modern deep learning models are usually pre-trained on large datasets like ImageNet before being fine-tuned to a specific task. This pre-training incurs huge compute costs, and thus the ML community relies on the large technology companies to provide the pre-trained models. This paper proposes a quicker and cheaper method of pre-training with distillation. This method involves using a sizable pre-trained teacher network to get a smaller student network up to speed, effectively pre-training it in a fraction of the time and cost. The numbers in this paper are impressive, and the researchers do a great job picking apart the different variables at play and how they influence the final performance. Some good science from the research team out of Baidu.
____
Think The Juice was worth the squeeze? Sign up here to receive it weekly.
